JP7369541B2 - Method of obtaining at least one important feature in a set of components of the same type and method of classifying such a set of components - Google Patents

Description

The present invention provides a method for obtaining at least one important characteristic in a series of components of the same type based on a data set obtained by a non-destructive testing method, and carrying out the evaluation of manufacturing processes and, in particular, the detection of trends, with respect to methods for classifying components, based on data sets obtained by non-destructive testing methods; The purpose is to be able to provide the following.

Methods that have been used to date to evaluate and control manufacturing processes through various process parameters using computed tomography data rely exclusively on information obtained from defects or discontinuities. These are observations compiled from individual evaluations, and therefore do not create relationships between a series of manufactured components (populations). In addition, it should be noted that most of the data obtained, especially in the case of computed tomography (CT), is not incorporated into the evaluation at all. These data are all areas within the component that are free of discontinuities and defects. In the field of CT, this is typically 90-95% of the data generated during inspection of each component investigated . In the case of non-defective components, these known methods do not yield any further information other than the information that the component is defective.

X-ray CT allows a three-dimensional representation of the components investigated using this method, especially of internal structures. The internal structure of a component can vary as a result of many process influences, and this result can be reflected in structural changes. These changes vary in nature depending on the manufacturing process and depending on the materials or combinations of materials used. Depending on the nature of the changes, these can be tolerated within a wide range but can be visualized with X-ray CT.

The object of the invention is therefore a method which makes it possible to train a system and subsequently inspect such a component, which during said inspection discontinuities leading to defects in the produced component are detected. The object of the present invention is to provide a method that can detect variations and, above all, trends in the manufacturing process at an early stage.

This object is achieved according to the invention by a method with the features of claims 1 and 5. Advantageous embodiments are specified in the dependent claims.

According to these, the object is a method for obtaining at least one important characteristic in a series of components of the same type on the basis of a data set obtained by a non-destructive testing method, The method is carried out in the form of computed tomography, and the method comprises :
- obtaining a 3D data set in each case for each component by inspecting a conditioned random sample of a plurality of components of known order of manufacture with said non-destructive testing method;
・Dividing a plurality of components in a random sample into good products and defective products;
・Extracting defect-free component regions from all of the component parts in the random sample;
- at least one feature from the extracted defect-free component region, the at least one feature being specific to the type of component and the manufacturing process for the component, and the at least one feature is specific to the component type and manufacturing process of the component ; machine learning approaches such as neural networks, multi-instance learning, statistical methods based on the field of data analysis , or determining using a statistical method based on a predetermined set of characteristics by genetic programming or conventional statistical methods;
- defining at least one feature and that characteristic as a trained classifier.

This method represents a first step designed as a learning phase. Its starting point is a conditioned random sample, ie a set of components whose manufactured order is known. The essential subdivision required for conditioning in the learning phase consists in dividing the random sample into good and bad products (OK/NOK). In the first step of the method, for example, machine learning methods are used. Machine learning methods automatically identify features that are specific to the process and type of component and that exhibit significant differences in their characteristics between good and defective products. Alternatively, the features may be selected from a predefined set.

In an advantageous development of the invention, in addition to the last step according to the invention (defining the at least one feature as a trained classifier), at least one process parameter is incorporated into the manufacturing of the component. and determining whether there is a correlation between the process parameter and at least one feature, and if there is a correlation, the at least one feature and its characteristics are defined as a trained classifier; In the case, the aforementioned correlation determination is made for different process parameters until there is a correlation, and then at least one feature is defined as a trained classifier, and if no correlation is found for any of the process parameters. It is provided that each step of claim 1 is performed for another feature, after which the steps in the present claim described above are repeated for the other feature. If the process parameters mentioned above are incorporated into the learning phase, there could in principle be the possibility of creating a direct correlation between the features identified in the volumetric data set and the process parameters. Therefore, control and automatic optimization of process parameters based on X-ray image information becomes possible.

In a further advantageous development of the invention, it is provided that a plurality of features are determined according to step d of claim 1 and that a correlation with at least one process parameter is generated by a combination of at least two features. In principle, a change in one process parameter is not only accompanied by a change in one characteristic, but is more complex, so that when one process parameter changes, multiple characteristics are affected. Additionally, a combination of features may increase the significance of the correlation.

For example, the following are used as process parameters: pressure, temperature, electrical and mechanical stresses or flows of the manufacturing process, ambient temperature or air humidity. It is ensured that these process parameters are (individually) assigned one-to-one to the component during its manufacture.

A further advantageous development of the invention provides that the method is carried out only for spatial regions of the component that can be predetermined. This reduces the scale of volumetric data considered for later steps, thereby speeding up the process.

According to claim 6, the object is a method for classifying a component in a series of components of the same type on the basis of a data set obtained by a non-destructive testing method, the method comprising : ,
i) whether a trend towards the characteristics of the characteristics of defective products is detected using the trained classifier obtained with reference to any one of claims 1 to 5; The analysis can be performed using machine learning approaches such as neural networks, multi-instance learning , statistical methods based on the field of data analysis, or statistics based on the characteristics of a predetermined set by genetic programming or traditional statistical methods. The steps are performed using a method of
ii) Communicate that no correction is required to the component manufacturing process if no such trend is detected , and/or if the process parameters are used in obtaining the classifier, performing automatic corrective action in the component manufacturing process by modifying the process parameters via an interface with a manufacturing machine. achieved by.

The second phase (after the learning phase) is to determine whether there is any process deterioration during the manufacturing process. Such process deterioration can lead to component deterioration and, in the worst case, component failure, but in some cases this deterioration can be avoided by changing the relevant process parameters. can be canceled, and in some cases this can be done automatically. The properties of this feature determined by the learning phase are used. Characteristics of the quality of the various components, their properties and possibly regions within the part are also identified in the learning phase, and in the second phase the application of the data learned in the first phase is carried out. (This second phase is therefore also referred to as the application phase). In the case of a series of tests, all components arranged for testing are investigated by means of computed tomography. The results in determining the characteristics are evaluated by trend analysis in conjunction with the process parameters that may be available. This trend analysis is based on machine learning or uses statistical methods, for example. Adjustment interventions in the process can be initiated if a trend towards developing the characteristics of a defective product (NOK) characteristic is detected. This intervention depends on the results of the trend analysis and can be performed automatically if the process parameters are part of the training data set and can be continuously detected during the application phase. According to the invention, therefore, it is possible to intervene in the process, thereby preventing the production of defective products. This allows prediction of process behavior with corresponding proactive reactions.

In an advantageous development of the invention, at least one defect-free component region is extracted in the investigated component and the method for classification described above is carried out only in this defect-free component region. provided. In such an approach, regions within the component are divided to increase the importance of features. This step in particular makes it possible to reduce the scale of the volumetric data considered in later steps and significantly reduces the time required for the evaluation.

In a further advantageous development of the invention, it is provided that the non-destructive testing method carried out is computed tomography, in particular X-ray tomography. A particular advantage of computed tomography is the high density of information obtained, which increases the likelihood of finding important features. Instead of X-ray tomography, it is also possible to use methods using ultrasound or millimeter waves, terahertz spectroscopy or thermography.

Further details and advantages of the invention are explained in more detail below with reference to examples of embodiments represented in the drawings.

FIG. 3 is a flow diagram of the learning phase according to the present invention. FIG. 3 is a flow diagram of the application phase according to the invention after performing the learning phase;

FIG. 1 represents the basic flow of the method in the learning phase according to the invention. This learning phase will be explained in more detail below.

To obtain the features, we first need a classified random sample (step number 1). That is, in the case of a simple two-classification problem, it is divided into, for example, a "good" class and a "bad" class. However, it is also possible to solve problems that work with more than two classifications. A random sample requires a sufficient number of elements. The number of elements depends on the complexity of the classification task; several hundred elements can produce good results. However, in individual cases millions of (partial) data sets may be required. Each element must be uniquely assigned to a classification so that the classification system can be trained.

In step number 2, defect-free component regions are extracted. Areas within the component can be excluded on the basis of quality criteria defined for the component under investigation in each case. For example, if porosity is critical to the quality of the component, component regions containing porosity are not used to obtain features. Other further quality criteria of the pore size can be, for example, geometric deviations from the nominal size or drawing specifications (CAD), cracks in the components, separation of the constituent materials, shrinkage holes, foreign object contamination.

Extraction of defect-free component regions can be performed manually, with an operator visually determining whether the extracted region is defect-free. Automatic extraction and evaluation is possible as well, since automatic image processing methods exist that are capable of detecting defects for all the aforementioned quality criteria. Therefore, a method for extracting defect-free component regions may be designed to step-by-step extract regions from a volumetric data set of components of regions and evaluate the regions using image processing methods. If a defect is detected, the region is discarded and not used for training. Regions that are not determined to be defective through image processing are included in training or application.

In step number 3, it is queried whether a predetermined feature selection is to be made.

If the answer to this is negative, then according to step number 4, the neural network or machine learning approach analyzes the inspection decisions made on the component and the internal structural characteristics of the component based on the classified random sample. offers the possibility to automatically determine features that show a correlation between In such approaches, models such as multi-instance learning, among others, are used to maximize learning outcomes.

In contrast, if the answer to the question of step 3 is affirmative, then according to step number 5, the predetermined feature is provided. To this end, conventional statistical methods from the field of data analysis may be used to identify features that produce the above-mentioned correlations. As a prerequisite, it is necessary that these features be selected in advance. This can be done manually by a trained image processing expert or through automated processes used for image processing (using genetic programming).

Then, according to step number 6 (following step number 4 or 5), a selection of important features and/or component areas is performed. This may be done via neural networks or data analysis, for example. In addition to the type of feature, the local density distribution, fiber orientation or local wall thickness variations, and the location within the component where the feature is characterized are also important for demonstrating the above-mentioned correlations. obtain. Restriction of one or more features to specific areas/regions of a component may contribute to the importance of the feature.

In step number 7, it is queried whether additional environmental data should be fed into the training of the classification system.

If the answer is yes, then in step number 8 the correlation between the process parameter and the characteristic of the feature is determined. This is similarly done using neural networks or data analysis. If a feature has been identified that ultimately allows an evaluation of the process quality, the task set in this step is to find a match between the specific process parameter and the characteristics of the feature. For example, analysis may determine that certain features take on certain values or form a pattern when the pressure is high. A correlation is thereby generated between the characteristics of the feature and one or more process parameters. If this feature were to be detected during operation, a pressure adjustment could be introduced as a countermeasure, for example.

Following step number 8, or if the answer to the query in step number 7 is negative, a trained classifier is established in step number 9. Once the learning process is complete, for example when a neural network is used, there will be a trained network (filter settings and weights for each neuron in the network). The network and its parameters represent a classifier. In the case of predetermined features, the trained classifier's knowledge represents the characteristics of the feature that are important for each classification (optionally also related to the position within the component where this feature is evaluated). ).

FIG. 2 represents the principle flow of the method in the application phase according to the invention. This application phase will be explained in more detail below.

If the important features, their specific characteristics and optionally their local relevance are selected through the method of the learning phase described above (step number 9), they are subjected to normal operations (a series of tests). ) used in Here, either all components (in-line operation) or a random sample of the relevant components (eg every 10th component) are investigated using X-ray CT according to step number 10.

Defect-free component regions are extracted in the same way as in training, according to step number 11.

Optionally, according to step number 12, the selection of relevant component regions is also performed based on the results of the learning phase.

Using the trained classifier previously established in the learning phase according to step 9, feature and classification calculations are then performed in step number 13. For this purpose, previously identified features are evaluated in potentially relevant component areas. The sum of the individual results (each feature within the various evaluated component regions) is then evaluated using a classification method (e.g. using a neural network) and, according to step number 14, a negative trend is determined. A determination is made as to whether or not it has been detected.

If no negative trend was detected in step number 14, the determination indicates to the operator that no remediation is necessary, according to step number 15.

If a negative trend is detected in step number 14, an inquiry is made according to step number 16 as to whether the process parameters have been learned.

In the simplest case, ie if the answer to this inquiry is negative, a warning is given about the process change according to step number 17. This means that the process tends to be more likely to produce defective products.

If the learned process parameters are present, the answer to the question of step number 16 is affirmative, and a correlation between the features and potential causes that may predict the occurrence of a discontinuity may be further indicated. . Such environmental data can be primary influencing factors such as manufacturing machine parameters (pressure, temperature, stress, flow), but secondary influences such as ambient temperature, air humidity and operator It can also be a factor. Crucial to the discovery of features during the learning phase is the detection of their importance over time in the manufacture of the component. Only through the time element can conclusions be drawn about the behavior of the process and the detection of trends, which ultimately becomes a prerequisite for proactive efforts to avoid defects.

If additional environmental parameters are involved, then according to step number 18, the method is used to automatically execute the process, and the identified process parameter (e.g., pressure) is changed via the classification system. . This counteracts negative trends in the manufacturing process and proactively prevents the formation of defective products.

Therefore, the invention can be summarized as follows:

The purpose of the method according to the invention is not to detect the exceedance of limits regarding unacceptable discontinuities, but to exploit variations between components that are within an acceptable range to control the process. be. This method requires a series of non-destructive tests, for which, in example embodiments, computed tomography is used. In this way, for each inspected component, a 3D data set is generated that reproduces the internal structure of the component. Through this series of inspections, changes in the manufacturing process can in principle be detected. This visualizes trends that are predictive in nature. Therefore, it becomes possible to predict process changes and thereby counteract their deterioration through intervention.

An essential part of this method is that information is obtained exclusively from discontinuities or defect-free component regions. Depending on component properties and process variations, there are various characteristics that can be correlated with component quality and process quality.

A crucial difference between the present invention and the prior art method is that the present method relies solely on information obtained primarily from defect-free areas of the component. However, it is also possible to use information from additional defect-prone areas to increase the importance of the characteristics of the feature . Methods that derive information from component defects do not allow meaningful information to be obtained from defect-free components in the sense of process evaluation and process impact. However, in the case of methods according to the prior art, a monitoring gap exists, since trends can become visible even in the case of non-defective components. Therefore, defect-free manufacturing is impossible because defects are always needed to obtain information. Therefore, such methods are exclusively reactive. In contrast, the inventive method described herein does not require defect-prone components to operate during the application phase, and thus is proactive. I can do it.

Claims (8)

A method for obtaining at least one important characteristic in a series of components of the same type based on a data set obtained by a non-destructive testing method, the method comprising:
said non-destructive testing method is carried out in the form of computed tomography;
The method includes:
a) obtaining a 3D data set in each case for each component by examining a conditioned random sample of a plurality of components of known order of manufacture with said non-destructive testing method;
b) separating the plurality of components in the random sample into non-defective products and defective products;
c) extracting defect-free component regions from all of the component parts in the random sample;
d) at least one feature from the extracted defect-free component region, the at least one feature being specific to a type of component and a manufacturing process for the component; Determining, using a machine learning approach such as a neural network or multi-instance learning, at least one feature that exhibits a significant difference in characteristics between good and defective parts over time in the manufacture of the part. and,
e) defining said at least one feature and its properties as a trained classifier. In addition to step e, at least one process parameter is incorporated into the manufacture of said component, and it is determined whether there is a correlation between said process parameter and said at least one characteristic and its properties; In some cases, said at least one feature and its properties are defined as a trained classifier, and if there is no correlation, said determination of correlation is performed for another process parameter until there is a correlation, and then If said at least one feature is defined as a trained classifier and no correlation is found for any of said process parameters, each step of claim 1 is performed on another feature, followed by said 2. A method according to claim 1, wherein the steps in this claim are repeated for said further feature. 3. A method according to claim 1 or 2, wherein a plurality of features are determined according to step d of claim 1, and a combination of at least two features produces a correlation with at least one process parameter. 4. The method according to claim 2 or 3, wherein the pressure, temperature, stress or flow of the manufacturing process, ambient temperature or air humidity are used as process parameters. Method according to any one of claims 1 to 4, wherein the method is carried out only for spatial regions of the component that can be predetermined. A method for classifying a component in a series of components of the same type based on a data set obtained by a non-destructive testing method, the method comprising:
The method includes:
i) using said trained classifier obtained with reference to any one of claims 1 to 5, whether a tendency developing towards the characteristic of defective products is detected; Trend analysis can be performed using machine learning approaches such as neural networks, multi-instance learning, statistical methods based on the field of data analysis, or statistical methods based on the characteristics of a predetermined set by genetic programming or traditional statistical methods. steps to be executed using
ii) if no such trend is detected, communicate that no correction is required to the manufacturing process of said component; if such trend is detected, communicate a warning; and /or if process parameters were used in obtaining said trained classifier, automatic corrective actions in the manufacturing process of said component by modifying said process parameters via an interface with a manufacturing machine; and performing the steps. 7. The method according to claim 6, wherein at least one defect-free component region is extracted in the inspected component and the method for classifying is performed only on the defect-free component region. A method according to any one of claims 1 to 7, wherein the computed tomography is an X-ray tomography.

Images